Who
Frontier
What
1,102 petaflop(s) per second
Where
United States (Oak Ridge)
When

The fastest computer is Frontier, installed at the Oak Ridge Leadership Computing Facility (part of the Oak Ridge National Laboratory) in Tennessee, USA. The computer has a High Performance Linpack (HPL) result of 1.102 exaflops (1,102 petaflops) according to the 59th edition of the TOP500 list of the world’s most powerful supercomputers, published in June 2022.

The industry-standard method used to measure the performance of a supercomputer is to see how fast it can complete the LINPACK Benchmark: a sequence of dense linear equations performed using double-precision floating-point arithmetic. The speed is given as the number of floating point operations (calculations) per second (flops). This can be most simply explained as a lot of calculations made using numbers that can be extremely small or extremely large.

Floating-point arithmetic is the method used in computing to store large numbers. A floating-point number consists of three parts: the "sign bit" (a single one or zero), which indicates whether the number is positive or negative; the "significand" (or "mantissa"), which is the significant digits of the number; and the "exponent", which indicates the position of the decimal point relative to the significant digits.

There are several ways to encode floating point numbers, but the most common are what are called "single precision" and "double-precision". In single-precision arithmetic, each number occupies 32-bits of memory (32 ones or zeros). First comes the sign bit, followed by an 8-bit exponent (allowing 256 possible values, or a range of -126 to +127), then a 23-bit significand (allowing for seven significant digits). Double-precision floating point numbers are encoded as 64 bits, allowing for an 11 bit exponent (-1023 to +1024) and a 52-bit significand (giving 15–17 significant digits). So, for example, 64,560,200 as a floating point number would be written as +6.45602e7.

Frontier can run the high-performance computing LINPACK benchmark at 1.102 exaflops (1,102 million billion floating-point operations per second), but has a theoretical peak performance of 1.685 exaflops, so this figure may improve as its systems are optimized. When performing mixed-precision calculations (involving half-, single- and double-precision floating-point numbers – which is all that many applications require) Frontier can run at a speed of 6.88 exaops (6,880 billion billion operations per second).

A typical desktop computer can run the LINPACK Benchmark at a speed of roughly 150–200 gigaflops (expensive gaming systems can go around 100 gigaflops faster), making them only 0.000027% as fast as Frontier. It would take a powerful desktop PC about 42 days to do what Frontier can do in a second.

Frontier weighs around 268 tonnes (296 US tons), occupies 680 m2 (7,300 sq ft) of floor space and uses up to 21 megawatts of electricity when running. It was built by Hewlett Packard's Cray supercomputer subsidiary and consists of 9,472 nodes, each one equipped with an AMD Epyc 64-core 2Ghz CPU, four Radeon Instinct GPUs, and 5 terabytes of flash memory. In order to keep all this cool, a bank of four 350-hp pumps circulate 100 gallons of cooled water through the system every second.